Tarik OS

01

The Problem

A strategy and innovation director who also hosts and produces a syndicated show and a food podcast, running work and personal life across two Google accounts, kept buying task managers like Sunsama and Akiflow and abandoning them within weeks. Each is a closed product with a fixed opinion about how you should work, so when it didn't fit there was nothing to modify, only a subscription to cancel.

02

The AI Design Decision

I made it voice first because I think faster out loud. Talking is what I've done professionally for twenty years, and a chat box makes you finish the thought before you can start typing it. Talking catches the half formed one. The decisions I'm prouder of are the two things I chose not to build. It remembers what I've said and can find it weeks later by meaning rather than by exact words, which is worth the extra machinery. I deliberately didn't build that for searching my own briefings, because the AI reads the titles well enough and a second search system would have been work with nothing to show for it. And I didn't let it decide its own plan. The morning briefing and the nightly tidy up follow a fixed list of steps on a timer, because for something I rely on at 7am, a boring thing that degrades gracefully beats a clever thing that might improvise.

03

The Architecture

1. 01Talk to it in the browser or call it on the phone
2. 02it decides what you actually want
3. 03it reaches for one of 30 tools, each of which can be switched off in a second
4. 04everything it learns goes in one place it can search later
5. 05it answers out loud, and every call is recorded so I can see what it did

04

The Responsible AI Constraint

Zola reads text I didn't write all day: emails, web pages, feeds. She can also act, drafting mail, writing to my calendar, driving a browser, calling my phone. So the danger is text in an email talking her into using one of those against me. Instead of trying to spot that text, I took the dangerous options away. The tool that phones me takes no phone number. The number lives on the server, so there's no argument to hijack. Nothing here filters bad input; there just isn't a lever for bad input to pull. And I removed each guard once, on purpose, to watch its test fail, because a safety test you've never seen fail is only a comment.

05

The Result

Two paid subscriptions cancelled. I'd been paying for Sunsama and Akiflow and using neither, and this replaced both within a week of starting it. It has run my day every day since August 7, and the morning briefing arrives on a timer whether or not I open the app. It currently holds 25 recorded conversations, 32 things it remembers about me, and 24 briefings it has written, across 26 tools that have done real work. The code is public.

06

If you remember one thing

I switched off each safety guard once, on purpose, to watch its test fail. A safety test you've never seen fail is just a comment.